An Information Criterion for Sensitivity Analyses of the Treatment Ignorability Assumption

• Activity Number: 569
• Type: Contributed
• Date/Time: Wednesday, August 3, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Chen-Pin Wang*+
• Companies: UTHSCSA
• Keywords: causal modeling ; treatment ignorability ; information criterion ; Kullback-Leibler Distance

Abstract

The treatment ignorability assumption (Rubin 1978) is critical in causal modeling. This paper proposes a Kullback-Leibler Distance based information criterion (KLD-IC) suitable for conducting sensitivity analyses for verifying the treatment ignorability assumption in the causal modeling framework. Here we focus on how the inference of certain statistic $T_{n}$ is affected by the violation of the treatment ignorability assumption. Stemmed from Goutis and Robert (1998), our proposed KLD-IC is the posterior mean of the KLD between the (predictive) distribution of a statistic $T_{n}$, under two likelihoods $r$ and $f$, where $r$ is the likelihood when the ignorability assumption is met and $f$ is the the likelihood when the ignorability assumption is violated. We examine the asymptotic properties of this KLD-IC under certain regularity conditions. We also show the impact of this asymptotic property for examples where only part of the regularity conditions are satisfied. We applied this KLD-IC to a study of pain among those with opioid dependence.